Impact of impurities in captured CO 2 on pipeline corrosion: Part II – Machine learning modelling

Abstract

The integrity of pipeline infrastructure in Carbon Capture and Storage systems is significantly compromised by corrosion induced by impurities in captured CO₂ streams. The review paper builds on the foundation developed in part-I by utilising an extensive literature-derived dataset to quantitatively assess the impact of key impurities on corrosion progression under varying thermodynamic conditions, including temperature, pressure, concentration, and exposure duration. Advanced data-driven modelling approaches, employing tree-based machine learning (ML) algorithms, facilitate a deeper understanding of multi-contaminant interactions and their influence on corrosion rates. The analysis identifies SO₂, H₂O, and O₂ as the most significant contributors to both general and localised corrosion processes. To systematically evaluate corrosion predictions, the dataset was divided into a 80% training set and a 20% testing set. Among the applied tree-based ML models – Decision Tree (DT), Random Forest (RF), and TreeNet (TN) – , the RF and TN regression algorithms demonstrated the lowest error rates in training. RF emerged as the most effective model for predicting corrosion rate values in new test data, while RF and TN exhibited comparable performance in classification tasks. However, given its lower computational complexity, RF may be the preferred choice for practical applications requiring efficiency and scalability.